Large language models generate or transform text and may support capabilities such as structured output, tool calls or other input types. This category includes model providers and services for running or accessing models. A desktop runtime, a hosted provider and an API gateway solve different parts of the same decision.
Begin with the task, data constraints and operating environment. Then choose a suitable model and access route. The five published members below are representative examples, not a ranking of model quality. Model names, versions and features change; check the current model documentation rather than assuming every model behind a familiar service behaves the same way.
| Tool | Access role | Decision to make |
|---|---|---|
| DeepSeek | A model-provider product family with chat and API entry points | Current model, documented features and the intended access route |
| Ollama | Run supported local models or use its cloud offerings | Local hardware versus cloud inference |
| LM Studio | Discover, manage and serve supported models on a computer | Runtime compatibility, model files and available memory |
| OpenRouter | Access a catalog of providers through an API interface | Model and provider selection, routing and pricing |
| Together AI | Hosted model inference and model-development services | Inference option, deployment needs and training scope |
DeepSeek's official site distinguishes its chat product and developer API. Select the appropriate route and inspect the current documentation before relying on a particular model capability. A chat subscription or interface should not be assumed to provide identical API access.
Ollama exposes models through its runtime and API, with local and cloud options. LM Studio offers model management, chat and local model-serving interfaces. These tools provide ways to run supported files; they do not make every downloaded model equally capable or suitable for the available machine.
OpenRouter's quickstart describes a unified route to its model catalog. Inspect model and provider details for the features you need. Together AI documents inference, fine-tuning and deployment services. Distinguish calling an existing model from training or operating a dedicated deployment.
Create a small representative set of anonymized requests with agreed labels, including ambiguous and out-of-scope examples. Define the required output schema and what should happen when no label fits. Run the same instructions and inputs through each candidate access route, recording the exact model configuration rather than just the platform name.
Check invalid outputs, incorrect classifications and requests needing human review. Measure response time and usage for your own workload, including realistic input length and repeated calls. If a local option is being considered, repeat the test on the intended hardware. Choose against the application's requirements and keep a record of the evaluation so a later model update can be checked.
Hosted API pricing and local operation have different cost structures. OpenRouter describes access and billing arrangements, while Together AI distinguishes inference and other services. Check DeepSeek's current API pricing through its official developer links. Do not infer API allowances from a consumer chat offer.
Local models need storage, memory and an appropriate runtime. Review the model's license separately from the application's terms. An interface described as compatible with another API can still differ in supported parameters, tool behavior or output handling. Include those differences in the integration rather than assuming that swapping a model identifier completes the migration.
No. It provides access and routing to models; the selected model and provider still determine important behavior.
Evaluate the actual task, latency, cost and hardware. Size alone does not establish suitability.
Keep the model identifier, provider, settings, prompt, evaluation cases and observed limitations so the decision can be revisited.